Real world results from KITTI-360 evaluation

[keonhee-han]

Hello,

Thank you for your work on the paper and your code publication. I wish to ask if there is already a pre-trained model for KITTI-360 or a pipeline setup for it. I have been currently working on setting up this dataset to see occupancy (density fields) results from the input view setup in the dataset. As I could see the experiment result for KITTI-360 in the paper Fig. 14 in Additional Qualitive results, I hope to see if the pre-trained model is available.

Would you share some insights regarding the training setup for KITTI-360? :)

Thank you!

[zubair-irshad]

Thanks for your question. We haven't released the pretrained model, but happy to help you reproduce our results on Kitti-360.

To summarize, our kitti-360 experiments were overfitting experiments i.e. trained on a single scene (not generalizable) as we mention in our supplementary, although depending on the use case, one could use our generalizable training scripts to train on multiple Kitti-360.

We train on Kitti-360 in an autoregressive manner (i.e. using last 3 source frames as input to the model) to predict the next frame. Here and here are our sample dataloader for kitti360 experiments on a single scene, which can be used with NeO 360's training script here. Hope it helps!